On Mon, May 09, 2011 at 07:39:46PM +0400, Max Brazhnikov wrote: > On Mon, 9 May 2011 15:41:05 +0300, Kostik Belousov wrote: > > You did not supplied enough information. > > Which of the processes is parent, which is child ? > > Note that there are other threads in the pid 18636. What does they do ? > > Here is backtraces from all threads > http://people.freebsd.org/~makc/automoc4.bt > 63373 is a parent now, 63374 is a child. > > There were no related changes in Qt4 and automoc4 sources, probably my update > from 8.2-PRERELEASE to STABLE a week ago triggered the issue.
It is obviously application bug, yes, I think my guess was right. Thou shalt not call non-async safe functions in thy child of multithreaded process. Since it is a race, I see it more curious that it did not manifested itself prevously. > > > If you would allow me to make some guess, then I could assume that pid > > 18640 is the child. Note that the child is waiting for the pthread > > mutex locked which protects the stdio' FILE structure. Now, assume > > additionally that the parent had the FILE locked in one thread while > > another thread did the fork. Then, the child process would never be able > > to obtain the lock because the lock was acquired by the thread that > > exists no longer (in the child process, only the thread that called > > fork is duplicated). > > > > In fact, I believe that you already reported a similar problem with > > malloc(3) some time ago. The root of the problem would be an undefined > > (and permitted by POSIX) behaviour of calling non-async signal safe > > functions in multithreaded process after fork. > > > > For malloc(3), this can be argued to be a quality of the implementation > > issue, but there is no reason to specially handle random mutexes, even > > from libc. If the mutex was locked during the fork time, the protected > > data structure is arguably in the inconsistent state after the fork in > > the child.
pgpxoq0JuHcPt.pgp
Description: PGP signature
